Why the AI cyber threat is rising

Added: 2026-05-26

[00:00:00]No pressure, but I think we have about 6 to 12 months to patch a gigantic chunk of software or face some really serious consequences. And the reason, of course, is this. The new AI models being released by companies like Anthropic, most notably Mythos. Sure, you heard of Mythos, right? It's the model that Anthropic said was so powerful, they couldn't actually release it. And honestly, you know, when I heard that, I was actually a bit skeptical. I wasn't sure they provided the evidence to really back that up. But since that moment happened about six weeks ago, new data has started to come out which honestly it shows that this stuff is extremely real. And that's what I want to talk to you about today. So let's start here. This is from Misilla, the makers of Firefox, the browser. And it shows their monthly bug fixes going back about a year and a half. And you can see for a long time Misilla have been fixing about roughly 30 bugs a month. Every month they do the same. And then in February they started something started really to be really exciting for them.

[00:00:58]They started to use AI. Look they got from here up to 76 that month. And then through anthropic they got access to mythos. And this was how many bug fixes they completed. This is the gap between what what we had before and Mythos. This is the jumping capability. And yeah, look, sure, it's just one company and also numbers of bug fixes. It's not always the best measure because you know, not all bug fixes are equal, right? Some can be quite small and some can be really big. But we're also getting independent assessments which roughly show the same thing. So this is a graph from the AI security institute based here in London. I think probably the world's leading AI security institute and it shows the results on their cyber assault course which simulates the takeover of a company.

[00:01:49]It's rather ominously called the last ones. I I don't know why.

[00:01:53]And you can see so back here Claude Sonnet 3.7 that was released in January 2025. And at that point AI could complete the initial reconnaissance of a network. It could get in and have a look around, snoop around, but that was really it. By the time Mythos came out, it was up to step 22 on average, and it could do privilege escalation. It could get inside the network and then work its way up, getting more and more privileges to attack the network. But now, this is the important point here, right? This was the moment of the mythos panic.

[00:02:26]Suddenly, everyone began to think about banks and national security. There was a big worry about it. But I want to show you what has happened since which is much less talked about.

[00:02:36]These are the results on the AI security institute test of the models that have been released since since mythos. You can see OpenAI released GPT 5.5 and GPT 5.5 cyber. Yeah, I know the the names in here are complete nightmare, but these are the OpenAI models. And also even more confusingly, Anthropic released a new version of Mythos. But the names aren't the point. The point is the results. Because the last one's test only has 32 steps. And on their best attempt, the best models can now take over the network. That simply wasn't possible before. And two points here, right? The first is that this isn't really about mythos. The OpenAI model does just as well here. This is about the overall progress of AI in cyber. And the second point, of course, as you can tell, this progress is moving really, really fast. So, of course, look, the people in AI are extremely extremely aware of this. They're taking steps to protect it. Anthropic have their project glass wing where they only give out mythos to certain select institutions like banks and open AI have a similar thing and there might be government versions in the future. that this is the problem right the US models you can keep tightly locked down but there's another set of models open weight or open source models these days mainly come out of China and as the name suggests they're open you can download them you can run them on your computer and they'll do whatever you want and people have been tracking the progress of these models you can see here these are results from uh Casey who are the um American version of the IC Security Institute and what they find consistently is that the the Chinese models which are on the orange line are always tracking the American models. They're just a bit behind.

[00:04:33]And in case you reckon the gap is about 8 months basically what that means is that 8 months after the American models achieve something, the Chinese models have it. And if the Chinese models have it, anyone can have it. And if anyone can have it, then cyber criminals can have it.

[00:04:49]So that's what we're dealing with here.

[00:04:50]The clock is ticking. We've got eight months or at least we think we do because there aren't really good measurements of this sort of thing. You know, we don't really have good tests of intelligence for this for this uh kind of stuff. So we think the gap might be 8 months, but it could be a lot smaller.

[00:05:09]These are measurements by the Stamford AI index. They reckon that the Chinese models, the open source models are roughly equivalent to the American models right now. So maybe we could be seeing uh open source models with myth with mythos level capabilities in two or 3 months.

[00:05:27]Like I say, we don't know for sure what's going on, but definitely the clock is ticking.

[00:05:33]And what happens when the clock stops?

[00:05:35]Well, I think I want to show you here how significant cyber attacks already are to the UK economy. They're a big drain. So, in 2025, it was estimated that cyber attacks cost the UK economy 14.7 billion. That's actually 0.5% of GDP.

[00:05:54]And to put that number in context, I've put it next to the expenditure of different government departments.

[00:05:58]Ministry of Justice, prisons, FCDO, that's diplomacy, defra farms, DCMS, the digital department. Cyber attacks already exceed uh or are comparable the expenditure of most of our major departments and this was preai. What happens if this doubles or triples? What happens if it's times by 10?

[00:06:22]And you know, one thing that I actually think that the the mythos panic slightly got wrong was that it was really focused on financial institutions and national security. And actually those are probably the organizations that with the help of AI can harden themselves to this kind of attack. I think it's much more likely that AI will end up attacking the soft underbelly of society. And really that kind of means all the services that we use every day. Hospitals, GPS, schools, all the services that you've probably heard about getting hit by ransomware, but maybe now they'll be hit even more frequently.

[00:07:00]And the good news is that we can defend against this stuff. It is possible to patch software, especially now we have AI. But are we going to are we going to respond with the speed we need to?

[00:07:11]Personally, I I think that AI will end up being amazing for cyber security. We will finally finally be able to have something we've never had ever before, which is truly secure software.

[00:07:21]But that's the long term. And you know, the short term matters too. And are we going to do enough to protect ourselves in that time frame?